Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Mistral releases one-trillion parameter model via guarded endpoint - OpenSmartRoute
Mistral AI released a new model designed to compete with both open and closed rivals. The company called Mistral Large 4 its latest large multimodal model. This release aims to leapfrog American and Chinese competitors in the field. French President Emmanuel Macron described this move as "a third way in AI." Europe remains a key player in the global artificial intelligence race. Mistral positions ML4 as an alternative to models that are either fully closed or often made in China.
The divide between closed and open models is growing wider every day. Closed models can be unplugged from their servers at any time. Open models are frequently developed by companies based in China. Mistral wants to offer a different path forward for its users. They call the new model ML4, short for Mistral Large 4. Nicknamed Le Chonk because of its massive size, it is definitely not small. However, it is not yet an open-weight model available for public download.
For now, users can only access the model through a specific endpoint. This endpoint includes public guardrails to manage safety and usage. Mistral plans to make the actual weights available in just three weeks. This timeline follows completion of rigorous safety testing procedures. The company wants to ensure the software is secure before sharing it widely. Pierre Stock, the VP of Science at Mistral, discussed these plans with TechCrunch.
Security concerns have been mounting recently among Mistral's core audience. Enterprises and institutions are worried about how their data might be used. An open-weight model is easier to audit than a closed one, according to Stock. This transparency helps organizations verify that the software behaves as expected. Mistral aims to prevent malicious use while allowing defense applications. They will work with trusted partners and governments during this phase.
The goal is to ensure open-source weights can be used to defend systems. They cannot be used to perform malicious attacks under current rules. Stock emphasized the importance of balancing innovation with safety standards. This approach addresses growing fears about AI being used for harm. Mistral hopes to build trust through careful governance and controlled access.
Another important behind-the-scenes aspect is how ML4 was trained entirely on Mistral's own compute. The company used only 4,000 NVIDIA GPUs for the entire training process. This number is two to three times less than their Chinese competitors. It is also significantly less than the closed source competitors in the market. Efficient training reduces costs and environmental impact compared to larger operations. Stock highlighted this efficiency as a key competitive advantage for Mistral.
With benchmark results still pending, Mistral hopes ML4 will be best in class among open-weight models. They especially want to lead outside of China's specific ecosystem. The company believes focused training can outperform closed models in specific areas. These areas are key to its customers' daily operations and goals. Multimodal capabilities can add significant value in these specialized fields.
According to Stock, ML4's optimized use cases include cybersecurity and finance. Chip design is another major area where the model excels. This capability is core to two of Mistral's main backers. Dutch giant ASML led its Series C funding round recently. Samsung led its Series D last month at a €21 billion valuation. That amount is about $24.39 billion in US dollars.
At the time, the company tried to convey that hosting Chinese models was not a pivot. They did not become a mere inference provider as a result. With Le Chonk in its corner, Mistral believes it should still be considered a frontier lab. The new model reinforces their position as an innovator rather than just a service provider. This distinction matters for their reputation and future partnerships globally.
AI, ai models, Europe, France, Mistral AI
When you purchase through links in our articles, we may earn a small commission. This doesn't affect our editorial independence.
Anna Heim is a writer and editorial consultant.
You can contact or verify outreach from Anna by emailing annatechcrunch [at] gmail.com.
As a freelance reporter at TechCrunch since 2021, she has covered a large range of startup-related topics including AI, fintech & insurtech, SaaS & pricing, and global venture capital trends.
As of May 2025, her reporting for TechCrunch focuses on Europe's most interesting startup stories.
Anna has moderated panels and conducted onstage interviews at industry events of all sizes, including major tech conferences such as TechCrunch Disrupt, 4YFN, South Summit, TNW Conference, VivaTech, and many more.
A former LATAM & Media Editor at The Next Web, startup founder and Sciences Po Paris alum, she's fluent in multiple languages, including French, English, Spanish and Brazilian Portuguese.
The Disrupt experience is meant to be shared. Get your pass and bring a colleague, partner, or peer at 50% off. Cover more ground by making connections, building momentum, and discovering what's next in the startup ecosystem.
At 19, founder raises $11M for Ghost, maker of a $3,499 computer for personal AI
Federal judge calls Flock 'indiscriminate mass surveillance'
Amazon responds to data center backlash, says it no longer uses NDAs
Google thinks SpaceX's Starship has to launch 1,800 times before space data centers get off the ground
World's first enhanced geothermal power plant completed in just 23 months
Google releases Gemini 4 Argon, called its most powerful model yet
The Pentagon taps Elon Musk and Palmer Luckey to help decide what the military should do next
Mistral Large 4 announcement - Mistral AI released a new model designed to compete with both open and closed rivals.
Mistral AI has officially announced the launch of Mistral Large 4. This is a new large multimodal model built for specific goals. The company aims to leapfrog both American and Chinese rivals in the market. French President Emmanuel Macron described this initiative as "a third way in AI." Europe remains an active participant in the global artificial intelligence race. Mistral positions ML4 as a strong alternative to existing models that are either fully closed or often made in China.
The divide between closed and open models is growing wider every day. Closed models can be unplugged from their servers at any time by the owner. Open models are frequently developed by companies based in China with different licensing terms. Mistral wants to offer a different path forward for its users and partners. They call the new model ML4, short for Mistral Large 4. Nicknamed Le Chonk because of its massive size, it is definitely not small. However, it is not yet an open-weight model available for public download.
For now, users can only access the model through a specific endpoint. This endpoint includes public guardrails to manage safety and usage restrictions. Mistral plans to make the actual weights available in just three weeks. This timeline follows completion of rigorous safety testing procedures. The company wants to ensure the software is secure before sharing it widely. Pierre Stock, the VP of Science at Mistral, discussed these plans with TechCrunch.
Security concerns have been mounting recently among Mistral's core audience. Enterprises and institutions are worried about how their data might be used by third parties. An open-weight model is easier to audit than a closed one, according to Stock. This transparency helps organizations verify that the software behaves as expected. Mistral aims to prevent malicious use while allowing defense applications. They will work with trusted partners and governments during this phase.
The goal is to ensure open-source weights can be used to defend systems. They cannot be used to perform malicious attacks under current rules. Stock emphasized the importance of balancing innovation with safety standards. This approach addresses growing fears about AI being used for harm. Mistral hopes to build trust through careful governance and controlled access.
Model specifications and training details - The model uses one trillion parameters and was trained on internal compute.
Mistral Large 4 is nicknamed Le Chonk in reference to its one trillion parameters. This number represents the total count of variables used in the neural network architecture. It is a massive scale compared to many other models currently available today. The model is designed to handle complex tasks across multiple input types simultaneously. These inputs can include text, images, and audio data streams.
ML4 was trained entirely on Mistral's own internal compute infrastructure. They did not rely on shared cloud resources or external partners for this work. The company used only 4,000 NVIDIA GPUs for the entire training process. This specific hardware count is a key detail about their operational efficiency. Stock noted that this number is two to three times less than their Chinese competitors. It is also significantly less than the closed source competitors in the market.
Efficient training reduces costs and environmental impact compared to larger operations. Mistral focused on optimizing every aspect of their training pipeline. They avoided unnecessary redundancy in their data processing steps. This strategy allowed them to achieve high performance with fewer resources. Stock highlighted this efficiency as a key competitive advantage for Mistral. It sets them apart from rivals who often overspend on hardware.
With benchmark results still pending, Mistral hopes ML4 will be best in class among open-weight models. They especially want to lead outside of China's specific ecosystem. The company believes focused training can outperform closed models in specific areas. These areas are key to its customers' daily operations and goals. Multimodal capabilities can add significant value in these specialized fields.
According to Stock, ML4's optimized use cases include cybersecurity and finance. Chip design is another major area where the model excels. This capability is core to two of Mistral's main backers. Dutch giant ASML led its Series C funding round recently. Samsung led its Series D last month at a €21 billion valuation. That amount is about $24.39 billion in US dollars.
At the time, the company tried to convey that hosting Chinese models was not a pivot. They did not become a mere inference provider as a result. With Le Chonk in its corner, Mistral believes it should still be considered a frontier lab. The new model reinforces their position as an innovator rather than just a service provider. This distinction matters for their reputation and future partnerships globally.
Access method and release timeline - Users can access the model now via an endpoint, with weights coming in three weeks.
Users can currently access Mistral Large 4 through a public guardrail endpoint. This is not a direct download link for the model weights themselves. The endpoint acts as a gateway to run queries against the trained system. Mistral plans to make the actual weights available in just three weeks. This timeline follows completion of rigorous safety testing procedures.
The company wants to ensure the software is secure before sharing it widely. Pierre Stock, the VP of Science at Mistral, discussed these plans with TechCrunch. He explained that they will work with trusted partners and governments during this phase. The goal is to prevent malicious use while allowing defense applications. They aim to balance openness with necessary security controls.
Security concerns have been mounting recently among Mistral's core audience. Enterprises and institutions are worried about how their data might be used by third parties. An open-weight model is easier to audit than a closed one, according to Stock. This transparency helps organizations verify that the software behaves as expected. Mistral aims to prevent malicious use while allowing defense applications.
The goal is to ensure open-source weights can be used to defend systems. They cannot be used to perform malicious attacks under current rules. Stock emphasized the importance of balancing innovation with safety standards. This approach addresses growing fears about AI being used for harm. Mistral hopes to build trust through careful governance and controlled access.
Another important behind-the-scenes aspect is that ML4 was trained entirely on Mistral's own compute. The company used only 4,000 NVIDIA GPUs for the entire training process. This number is two to three times less than their Chinese competitors. It is also significantly less than the closed source competitors in the market. Efficient training reduces costs and environmental impact compared to larger operations.
With benchmark results still pending, Mistral hopes ML4 will be best in class among open-weight models. They especially want to lead outside of China's specific ecosystem. The company believes focused training can outperform closed models in specific areas. These areas are key to its customers' daily operations and goals. Multimodal capabilities can add significant value in these specialized fields.
According to Stock, ML4's optimized use cases include cybersecurity and finance. Chip design is another major area where the model excels. This capability is core to two of Mistral's main backers. Dutch giant ASML led its Series C funding round recently. Samsung led its Series D last month at a €21 billion valuation. That amount is about $24.39 billion in US dollars.
At the time, the company tried to convey that hosting Chinese models was not a pivot. They did not become a mere inference provider as a result. With Le Chonk in its corner, Mistral believes it should still be considered a frontier lab. The new model reinforces their position as an innovator rather than just a service provider. This distinction matters for their reputation and future partnerships globally.
Security and governance approach - Mistral aims to prevent malicious use while allowing defense applications.
Security concerns have been mounting recently among Mistral's core audience. Enterprises and institutions are worried about how their data might be used by third parties. An open-weight model is easier to audit than a closed one, according to Stock. This transparency helps organizations verify that the software behaves as expected. Mistral aims to prevent malicious use while allowing defense applications.
The goal is to ensure open-source weights can be used to defend systems. They cannot be used to perform malicious attacks under current rules. Stock emphasized the importance of balancing innovation with safety standards. This approach addresses growing fears about AI being used for harm. Mistral hopes to build trust through careful governance and controlled access.
Mistral will work with trusted partners and governments during this phase. They want to ensure the open-source weights can be used to defend systems. They cannot be used to perform malicious attacks under current rules. Stock emphasized the importance of balancing innovation with safety standards. This approach addresses growing fears about AI being used for harm.
Training efficiency compared to competitors - The model used significantly fewer GPUs than Chinese or closed rivals.
Another important behind-the-scenes aspect is that ML4 was trained entirely on Mistral's own compute. The company used only 4,000 NVIDIA GPUs for the entire training process. This number is two to three times less than their Chinese competitors. It is also significantly less than the closed source competitors in the market.
Efficient training reduces costs and environmental impact compared to larger operations. Mistral focused on optimizing every aspect of their training pipeline. They avoided unnecessary redundancy in their data processing steps. This strategy allowed them to achieve high performance with fewer resources. Stock highlighted this efficiency as a key competitive advantage for Mistral. It sets them apart from rivals who often overspend on hardware.
With benchmark results still pending, Mistral hopes ML4 will be best in class among open-weight models. They especially want to lead outside of China's specific ecosystem. The company believes focused training can outperform closed models in specific areas. These areas are key to its customers' daily operations and goals. Multimodal capabilities can add significant value in these specialized fields.
According to Stock, ML4's optimized use cases include cybersecurity and finance. Chip design is another major area where the model excels. This capability is core to two of Mistral's main backers. Dutch giant ASML led its Series C funding round recently. Samsung led its Series D last month at a €21 billion valuation. That amount is about $24.39 billion in US dollars.
At the time, the company tried to convey that hosting Chinese models was not a pivot. They did not become a mere inference provider as a result. With Le Chonk in its corner, Mistral believes it should still be considered a frontier lab. The new model reinforces their position as an innovator rather than just a service provider. This distinction matters for their reputation and future partnerships globally.
Why it matters
Mistral Large 4 represents a significant step forward for European AI development. It challenges the dominance of American and Chinese models in the global market. The efficiency gains could lower barriers to entry for other organizations. Companies might spend less on hardware while achieving similar results. This shift could accelerate innovation across various industries worldwide.
The security-focused release strategy addresses growing concerns about AI misuse. Organizations can audit the model more easily than with closed alternatives. This transparency builds trust among enterprises and institutions globally. It sets a new standard for responsible AI development and deployment. The approach balances openness with necessary safety controls effectively.
What to do
Engineers should monitor the public guardrail endpoint for immediate access. They can test the model's capabilities in cybersecurity and finance scenarios. Managers should prepare to evaluate the weights when they become available in three weeks. Compare ML4 against other models using relevant benchmarks and use cases. Check if the efficiency gains align with your budget and infrastructure needs.
